An efficient computational offloading method using deep reinforcement learning in edge-end-cloud

Abstract

Due to certain limitations of cloud and edge computing, the issue of delayed response arises from both. We propose an edge-end-cloud computational unloading solution based on deep reinforcement learning. Firstly, we introduce the pre-division algorithm to facilitate the implementation of the second stage and address the threshold selection of the calculation unloading strategy. Secondly, we analyze the computing resources within the DQN and Q-learning frameworks. Finally, we present the parameter verification of the blockchain. The integration of blockchain technology enhances the security and credibility of data transmission. Additionally, blockchain technology can strengthen the credibility of the edge ecology and mitigate the single point of trust risk encountered by traditional centralized architectures on the edge side. The experimental results indicate that the proposed computational unloading strategy in this paper decreases the edge-end-cloud architecture using DQN and Q-learning by approximately 40% compared to other computing strategies. When we adjust the server’s computing power to $F=10$ GHz/s, the energy consumption of Q-learning and DQN becomes nearly identical, suggesting that if the server’s computational power is sufficiently strong, the unloading results can often be more favorable.